nx_stl.nx

nx_stl.nx -- binary STL mesh reader (1980s 3D Systems format, public domain). Per cardinal NISHI_3D_PRINT_ROADMAP scope: STL spec is public domain since 1980s; reading it is not a license entanglement (Cardinal 0). Composes nx_le + nx_fp32_q14 + nx_hash + nx_mesh. Binary STL on-disk layout: bytes 0..79 header (ignored; ASCII STL detection: starts with "solid"; we refuse ASCII for P0.2a) bytes 80..83 little-endian uint32 = n_tris bytes 84.. per-triangle records, 50 bytes each: 0..11 normal vector (3 × f32) IGNORED -- we trust mesh winding, normals are often wrong in user-exported STL 12..47 3 vertices × 3 × f32 = 9 floats 48..49 attribute byte count (IGNORED) Total file size = 84 + 50 * n_tris. Vertex dedup strategy: Raw STL has 3 verts per tri with massive duplication (closed mesh of n tris has ~n/2 unique vertices, ~50% saving). We hash each (x_q14, y_q14, z_q14) triple via nx_hash_fnv1a_bytes into nx_hash; collisions are resolved by direct (x,y,z) verification against the existing mesh vertex. False matches from hash collisions allocate a new vertex (slight over-alloc; correctness preserved). Capacity heuristic: Worst case (no dedup): 3 * n_tris vertices. We size nx_hash to the next power of 2 above n_tris * 4 to keep load factor < 0.75. Manifold-check expectation: a well-formed closed STL produces a mesh that nx_mesh_is_manifold accepts. Real-world STL from scanners (Christus from Scan The World) sometimes has minor non-manifold artefacts (T-junctions, near-duplicate verts at quantization boundaries); those are the slicer's problem, not the parser's.